Efficiency and Adaptability of Deep Learning Architectures: A Performance Comparison of CNN, Transformer, and Hybrid Models

Innovations in deep learning architectures are driving a paradigm shift in computer vision. While traditional convolutional neural networks (CNNs) have long dominated image tasks due to their local perceptual advantages, the Transformer has rapidly penetrated the vision domain after emerging in natural language processing through its global attention mechanism. However, the high computational complexity and data dependency of the Transformer limit its practical applications. This study systematically compares the differences in computational efficiency, number of parameters, and task adaptability among CNN (ResNet-50), Transformer (ViT-B/16), and hybrid architecture (ConvNeXt-T) in the context of architecture evolution. The experimental results show that ConvNeXt-T architecture exhibits excellent performance on several standard datasets. The hybrid architecture demonstrates significant advantages in edge device deployment scenarios by fusing local convolution with global attention. The findings of the study offer empirical support for the selection of models in resource-limited settings and indicate pathways for the optimization of multimodal models.